Meituan ships LongCat-2.5, a 1.6T-parameter (about 48B active) native multimodal agentic model, with API pricing of $0.30/$1.20 per 1M tokens

Background and the problem

Meituan's LongCat team released LongCat-2.5-Preview on September 25 and announced the full LongCat-2.5 in its API platform change log on November 2. As of writing, the official @Meituan_LongCat X account has not posted about it, so this story relies on the official technical blog, change log and pricing page; the X post cited comes from the tracker account AI Tracker Bot. Long-horizon agent work such as full-project builds, large codebase changes, and browser and spreadsheet tasks is where frontier models now compete, and LongCat-2.5 targets both that gap and the constraint of training on domestic compute.

How it works

According to the blog, LongCat-2.5 is a MoE with 1.6 trillion total parameters and about 48 billion active per token, trained further from LongCat-2.0. Visual signals become multi-level discrete tokens (the RVQ approach from LongCat-Next) that share the same backbone as text, with no separate vision encoder in the training graph. Training adds more long-horizon trajectories and large-scale long-horizon RL, turns real agent failure trajectories into verifiable training tasks, and mixes training across harnesses. Long-context training and W4A4 inference run on domestic AI accelerators with under 80 GB of memory per card.

Benchmarks and numbers

All figures are vendor-reported: Terminal Bench v2.1 88.4 (Claude Opus 5 89.1, GPT-5.6-Sol 88.8, DeepSeek-V4.1-Flash 90.6), Terminal Bench v4.0 38.5 (GPT-5.6-Sol 37.3, Claude Opus 5 53.9), DeepSWE v1.1 64.6, BrowseComp 91.9, MCP Atlas 83.1, HLE with tools 63.1 and Vision2Web 78.6. The blog says rival scores marked with an asterisk come from public reports or leaderboards and the rest were run by Meituan and are indicative only; coding scores such as Terminal Bench were run in the Claude Code harness. No third party has reproduced them yet.

Getting started

It is API-only for now (pay-as-you-go and token packs). The model name is LongCat-2.5, and the API supports both the OpenAI format (https://api.longcat.chat/openai) and the Anthropic format (https://api.longcat.chat/anthropic), so it plugs into Claude Code, Codex, Hermes Agent, OpenCode, Cline and similar tools, each with its own setup page in the docs. The limited-time price is $0.30 per 1M uncached input tokens, $0.006 cached input and $1.20 output, and Meituan notes prices may change. Run the same tasks on your own codebase against your current model before switching.